“Beyond Translation: Nature-Driven Deep Learning Architectures for Biological Analysis”
Recent advances in deep learning have revolutionized many fields, yet their application to biological systems often overlooks crucial domain-specific constraints. We present a novel framework that fundamentally integrates biological invariances into deep learning architectures, moving beyond the simple translation of popular models to biological applications. Our approach encompasses two major directions: First, we develop a new generation method for antimicrobial peptides that incorporates physicochemical properties and structural constraints directly into the model architecture. Second, as part of the Immucan consortium, we introduce innovative deep learning methods for analyzing Imaging Mass Cytometry (IMC) data, enabling sophisticated spatial analysis of tissue microenvironments. By prioritizing biological principles in our model design, we achieve both improved performance and better interpretability compared to conventional approaches. Our framework demonstrates that incorporating domain knowledge is crucial for developing reliable and effective deep learning solutions in life sciences.
Bio
Marcin Możejko is a PhD student in Prof. Ewa Szczurek’s laboratory at the University of Warsaw, where he focuses on applying generative models to biological entities. With over 12 years of experience spanning both industry and academia, he brings a unique perspective to computational biology research. His professional journey includes roles at leading organizations such as Microsoft, PwC, TCL Research Europe, and Sigmoidal, providing him with extensive expertise in both theoretical and applied machine learning. In his current research, Marcin collaborates with prestigious institutions worldwide, including the Bodenmiller lab at the University of Zurich, the Feinberg School of Medicine at Northwestern University in Chicago, and Helmholtz AI in Munich. He is also an active contributor to the Immucan consortium, where he develops novel deep learning methods for spatial biology analysis.